Anthropic Files for IPO, Testing Public Appetite at 965B

The confidential S-1 caps a historic fundraising month and signals that private AI valuations have reached their ceiling.

June 5, 2026 | Reading time: 9 minutes | Issue #180

Anthropic confidentially submitted a draft registration statement on Form S-1 to the SEC on June 1, the company announced in a brief statement. The filing gives Anthropic the option to go public after the SEC completes its review, but the company explicitly cautioned that the proposed initial public offering will depend on market conditions and other factors. No share count or price range has been set.

The filing arrives three days after Anthropic closed a $65 billion Series H at a $965 billion post-money valuation, making it one of the most valuable private companies in history. The round was led by crossover investors including Coatue and Tiger Global, with participation from sovereign wealth funds in Singapore and the Middle East. In the same week, Google raised $22.5 billion in debt explicitly earmarked for AI infrastructure, and SpaceX closed a $12.5 billion secondary transaction. The concentrated burst of capital suggests the private market is attempting to front-run public-market appetite before liquidity windows narrow.

An IPO would subject Anthropic to quarterly revenue disclosure, shareholder litigation exposure, and the governance constraints of a public board. That structure conflicts with the company's public benefit corporation charter, which grants its Long-Term Benefit Trust an effective veto over major strategic decisions. How Anthropic reconciles fiduciary duties to public shareholders with its existing governance architecture is an open question, and one that the SEC review process will scrutinize closely. The bet among crossover investors is that public-market multiples for frontier AI will exceed private-market comparables. The risk is that they are wrong, and that a public Anthropic trades like an infrastructure stock with software margins.

Apple will relaunch Siri at WWDC — still labeled beta

Bloomberg reported on June 5 that Apple will unveil an overhauled Siri and a standalone Siri app at its Worldwide Developers Conference next week. The new assistant is expected to feature deeper AI integration, on-device reasoning, and tighter app control. But internally, Apple is still labeling the software "beta" and "preview," and Mark Gurman reports that the company may implement a waitlist for access, similar to the staggered rollout of Apple Intelligence in 2024. The framing suggests Apple is managing expectations after the delayed and underwhelming launch of its first-generation AI suite. A Siri that requires a waitlist nine months after announcement is not a product ready for mass adoption.

OpenAI rolls out a new memory architecture

OpenAI published a research post on June 4 describing a new memory synthesis system for ChatGPT called Dreaming. The update replaces the saved-memories feature launched in April 2024 with a more scalable architecture that optimizes for freshness, continuity, and relevance across multi-year time horizons. The system is rolling out to Plus and Pro users in the US this week. Memory is a second-order feature — it does not change what the model knows, but changes how it relates to individual users over time. The practical effect is that ChatGPT will retain project context, preference patterns, and conversational history with less staleness. For OpenAI, memory is a retention mechanism: the more context a user has invested in the system, the higher the switching cost to a competitor.

NVIDIA and Microsoft build a unified agentic stack

At Microsoft Build on June 2, NVIDIA and Microsoft announced an expanded partnership to deliver a unified hardware-software stack for agentic AI across Windows devices, Azure cloud, and local deployments. The announcement includes RTX Spark laptops with 1 petaflop of AI performance, DGX Station for Windows deskside supercomputers powered by the GB300 Grace Blackwell Ultra chip, and NVIDIA OpenShell — a secure runtime integrated into GitHub Copilot. The stack also brings NVIDIA GPU-accelerated Microsoft Fabric and NVIDIA open models to Microsoft Foundry. The through-line is that Microsoft is positioning Windows as the default host environment for enterprise agents, while NVIDIA supplies the silicon. For developers, the pitch is write once, deploy across cloud, local, and edge without retuning.

Anthropic maps a year of AI-enabled cyber attacks

Anthropic published a detailed analysis on June 3 of 832 accounts banned for malicious cyber activity between March 2025 and March 2026. The study, conducted with MITRE ATT&CK framework mapping, found that 67.3% of banned actors used AI to write malware, and 6.5% used it for lateral movement inside compromised networks. More significantly, the share of actors classified as medium-risk or higher jumped from 33% in the first half of the study period to 49% in the second half — a shift Anthropic attributes to AI making less sophisticated attackers more capable. The firm also concluded that the MITRE ATT&CK framework does not fully capture AI-enabled attack chains, suggesting that industry-standard threat taxonomies are lagging behind the technology they are meant to describe.

Compute Watch

NVIDIA's June 2 partnership with Microsoft is not the only compute story this week. On June 4, NVIDIA published a long-form overview of its South Korea ecosystem, detailing sovereign AI infrastructure partnerships with LG, SK Telecom, and Korean government research labs. The company is positioning itself as the default vendor for nations that want domestic AI capability without building fabs — a category that now includes India, Korea, France, and Saudi Arabia.

The RTX Spark and DGX Station for Windows announcements matter because they extend NVIDIA's reach from data-center bulk purchasing to individual knowledge-worker desktops. A DGX Station with 748GB of coherent memory and 20 petaflops of FP4 performance can run frontier models up to 1 trillion parameters locally. That changes the unit economics for enterprise AI: instead of renting API tokens, firms can amortize inference across a hardware depreciation schedule. The question is whether IT departments will accept the security and maintenance burden of on-premise frontier models, or whether the convenience of cloud APIs still wins.

India Lens

Rest of World published a feature on June 5 documenting a reverse brain drain in Indian AI talent. Every month, two to three Indian-origin researchers in Silicon Valley contact Mumbai-based AI venture fund Activate asking how to return and work in India's AI sector. The shift is driven by three factors: US big-tech layoffs between 2023 and 2025, H-1B visa unpredictability under the Trump administration, and the maturation of Indian AI startups beyond back-office implementation roles.

The supply-demand imbalance is severe. Neeti Sharma, CEO of staffing firm TeamLease Digital, told Rest of World that India has only one qualified engineer for every 10 open GenAI positions. OpenAI and Anthropic have both begun expanding engineering presence in India — Anthropic opened a Bengaluru office in May, and OpenAI has been hiring technical roles in the country since early 2026. Early-stage Indian AI firms pay 50-75% of US big-tech salaries, but compensate with stock options that could close the gap if any of these companies reach scale. Anuj Agrawal, founder of recruitment firm Zyoin Group, put it directly: "The aura hasn't disappeared, but the calculus has fundamentally shifted."

Eastern Front

DeepSeek topped the June trending-software index published by Ramp on June 4, as reported by the South China Morning Post. US firms are sending data directly to DeepSeek's China-hosted servers rather than self-hosting the open-source weights, according to Ramp economist Ara Kharazian. DeepSeek's V4 Pro is priced at roughly one-quarter of its US equivalents after a permanent 75% reduction announced in May. The pattern extends beyond DeepSeek: Fireworks AI, Fal AI, and DeepInfra also ranked among June's trending vendors, suggesting open-weight inference platforms are reaching procurement parity with closed APIs on enterprise benchmarks.

Meanwhile, Zhipu AI's GLM-5 reached open-source SOTA on SWE-bench Verified, matching Claude Opus 4.5, and Baichuan continues to push its medical AI vertical with the M3-Plus model offered free to healthcare service partners under its "海纳百川" program. Chinese labs are not just competing on model capability — they are competing on price, domain specificity, and outbound enterprise adoption.

The View

Anthropic's S-1, DeepSeek's Ramp index victory, and India's talent reverse-migration are three data points describing the same structural shift: the AI industry is splitting into three zones with incompatible economics.

Zone one is the public-market frontier, where Anthropic is about to test whether $965 billion in private valuation holds up under GAAP accounting and quarterly earnings calls. Zone two is the Chinese export model, where DeepSeek and Zhipu compete not on absolute capability but on intelligence-per-dollar, and are winning procurement decisions from cost-conscious US firms. Zone three is the sovereign-compute middle ground, where India, Korea, and the Gulf states are building domestic infrastructure not to train frontier models, but to deploy them under local governance rules.

These three zones do not converge. They diverge. The US is betting that capital markets and distribution lock-in will preserve margin. China is betting that price and open weights will erode that margin before lock-in sets in. The rest of the world is betting that owning the compute layer — even if it means renting the model — is the only hedge against dependence on either superpower. What unites all three is pressure on the middle: the integration layer between model and enterprise. That is where the next set of winners will emerge, and where the current incumbents are most exposed.

The Miss

Rest of World published an essay on June 4 arguing that compute scarcity is becoming the primary driver of AI innovation outside Silicon Valley. The piece, by Farcana CEO Ilman Shazhaev, notes that data centers consumed 1.5% of global electricity in 2024, that chip access is increasingly jurisdiction-dependent, and that builders in the Global South are designing around constraints rather than adding capacity. The argument has immediate policy relevance: if the US, China, and Europe monopolize both chip fabrication and model training, the rest of the world will be forced into a pure consumption role — or will build alternative stacks. The essay received limited attention relative to its structural implications. Coverage to date: one outlet.

Pull Quotes

"The aura hasn't disappeared, but the calculus has fundamentally shifted." — Anuj Agrawal, Zyoin Group

"In probably the biggest sign that companies are looking for cheaper alternatives to OpenAI and Anthropic, some are willing to use cheaper, Chinese models, sending US data back and forth from China-hosted servers." — Ara Kharazian, Ramp Economics Lab

"The proposed initial public offering will depend on market conditions and other factors." — Anthropic, SEC filing

Reads & Links

• Anthropic confidentially submits draft S-1 to the SEC
• Bloomberg: What to expect at WWDC 2026
• OpenAI: Dreaming memory update
• NVIDIA partners with Microsoft on unified agentic AI stack
• Anthropic: AI-enabled cyber threats report
• SCMP: More US firms turn to China's DeepSeek
• Rest of World: Silicon Valley's lure fading for India's tech talent
• NVIDIA: Seoul Purpose — Korea ecosystem
• Rest of World: Scarcity driving AI innovation outside Silicon Valley

Out

The question is whether public markets will value Anthropic like a software company or like an infrastructure play — and whether the answer matters for the rest of the stack.

By Neo